Эксперимент: Dissipation-Budgeted Nonreversible Sampling (#1336)

{ "worked": true, "confidence": 8, "verdict": "Built a reproducible 1D double-well Langevin MVP with constant nonreversible drift, cumulative quadratic-dissipation budgeting, first-passage detection, survival curves, and hazard estimates. The survival identity was exact numerically, and the dissipation formula matched to 0.0031 absolute error, approximately one timestep. Uncapped drift produced a 45.7x tail-hazard speedup at mean Q=7.26; under a Q=0.5 budget, hazard improved 1.13x, with log hazard gain versus Q showing an approximately linear fit (R²=0.994). This is a promising small-system signal, but not evidence of a general ML sampling win.", "metrics": { "baseline": "Tail hazard 0.00807; hit fraction 0.0701; mean hit time 7.33.", "idea": "Uncapped u=0.5: tail hazard 0.3688 (45.7x), hit fraction 0.973, mean hit time 4.65, mean Q=7.26. Budgeted u=0.5 with Q<=0.5: tail hazard 0.00910 (1.13x), hit fraction 0.0842, mean hit time 7.05, Q=0.50. Log hazard ratio versus Q fit R²=0.994, slope C=0.346." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 experiment.py", "files": [ "experiment.py", "results.json" ], "limitations": "Only a 1D analytic double-well and constant drift were tested; no learned score model, MLP control, multidimensional nonreversible rotational drift, terminal KL, FLOP comparison, or independent repeated-seed confidence intervals were evaluated. The budget controller exhausts its budget early, so tight-budget improvement is modest." }